Weather Prediction on Mars as a Multivariate Time Series Forecasting Problem

Accurate weather prediction on Mars is imperative for the safety of future human explorers and maximising the scientific return from robotic missions. Conventional physics-based numerical weather prediction models face challenges due to sparse observational data and the intricate Martian atmosphere. In this paper, we propose to use Machine Learning (ML) to forecasts Martian weather using the OpenMARS dataset and compare it to a physics-based numerical weather prediction model. OpenMARS is a reanalysis dataset that merges spacecraft observations and a Mars Global Circulation Model covering over a decade of Mars years, including three large dust storm events. We employ multiple ML models for time series forecasting to evaluate their performance against the OpenMARS dataset. The dataset includes variables such as surface pressure, temperature, near-surface winds, dust column, and water vapour, with a temporal resolution of two hours local time. Focusing on a 1-dimensional time series at a specific landing site location, resembling conditions for human exploration, we systematically train and test various ML models. Multiple ML models are efficient in the prediction of the dynamical variables up to one day in the future, with the TCN and TiDE models particularly effective at reproducing realistic intrinsic variability, but predicting the onset of a dust storm event remains challenging. Our findings contribute insights into Martian weather prediction, emphasising the potential and limitations of current ML-based approaches for timely decision making in future Martian missions. A replication package, including the OpenMARS dataset and the benchmarking results are publicly accessible at https://github.com/amelBennaceur/OpenMarsML. We hope that this work encourages collaborative efforts and advancements in ML for Martian weather research.